引入自上向下特征融合的小目标检测算法

针对原始SSD(Single Shot Multibox Detector)算法未充分利用各特征层之间关系导致浅层特征层缺乏小目标语义信息的问题，为了提高对小目标的检测能力，提出了一种结合PANet多尺度特征融合网络和自上向下特征融合路径的TTB-SSD(Top to Bottom SSD)改进算法。首先，使用PANet多尺度特征融合网络对特征进行反复提取，从而获得丰富的多尺度语义信息；然后，使用一种深层特征融合模块将浅层特征层的空间信息传递到深层特征层，进而更准确地对小目标进行定位；最后，为了增强浅层特征层的语义信息，构造了自上向下的特征融合路径，从而强化浅层对小目标检测的准确率。实验结果表明，在PASCAL VOC2007测试集检测的mAP(Mean Average Precision)值达到80.5%，对目标的mAP较原始SSD提高了5.7%，证明了该算法对小目标检测的有效性。

A Small Object Detection Algorithm with Top-to-Bottom Feature Fusion

For the problem that the original Single Shot Multibox Detector(SSD) algorithm does not make full use of the relationship between the feature layers,in order to improve the detection ability of small objects,an improved Top to Bottom SSD(TTB-SSD) algorithm combining PANet multi-scale feature fusion network and top-down feature fusion path is proposed.First,the PANet multi-scale feature fusion network is used to repeatedly extract features to obtain rich multi-scale semantic information.Then,a deep feature fusion module is used to transfer the spatial information of the shallow feature layer to the deep feature layer,so as to locate the small target more accurately.Finally,in order to enhance the semantic information of the shallow feature layer,a top-to-bottom feature fusion path is constructed to enhance the accuracy of small target detection in the shallow layer.The experimental results show that the mean average precision(mAP) value detected in the PASCAL VOC2007 test set reaches 80.5%,and the mAP of the target is improved by 5.7% compared with the original SSD,which proves the effectiveness of the algorithm for small target detection.